Anthropic Discloses Quarterly Revenue and Enters Accelerated Scaling Tier, Tencent Releases Hyra 2.0 to Upgrade Scientific Research Agents

This issue focuses on the financial and benchmark progress of frontier large model enterprises, as well as the latest industry developments in agent architecture and infrastructure.

The essentials

Anthropic Reports Quarterly Financial Data and Faces Accelerated Scaling Thresholds

Anthropic's submitted revised prospectus shows third-quarter revenue of nearly 17 billion dollars, an increase of over 1100% year-over-year, while quarterly net loss reached 23.3 billion dollars. Internal estimates indicate its future models may soon reach the significant accelerated progress tier in its responsible scaling policy, with large-scale infrastructure contract commitments.

Tencent Hunyuan Releases Hyra 2.0 to Advance Automated Scientific Research Agents

Tencent Hunyuan released Hyra 2.0, shifting automated research priorities toward agent-driven research and upgrading scientific models and tools. The system adopts a 3-layer division of labor and integrates into actual R&D environments such as WeChat Work and code repositories, with related achievements open-sourced on GitHub.

Source ↗

MIT Paper Proposes Weight Space Programming to Significantly Improve Small Model Performance

An MIT CSAIL paper shows large models can natively read and write other neural network weights. Experiments edited 0.016% of the weights of Qwen3.5-0.8B, fixing the issue of getting stuck in repetitive reasoning and improving its math score by 6 times.

In brief

SemiAnalysis Releases GDPx Benchmark

Among the released agentic capability frontier benchmarks, the current best model score is 12%, which Google's CEO reposted and praised.

Source ↗

OpenAI Releases LifeVal to Evaluate Daily Tasks

Measures model performance on ordinary human daily tasks such as finding childcare and managing personal finances.

Anthropic Launches Message Threads Public Beta

Conversation cache is stored on the server side, and the client sends only incremental messages each round, shortening the time-to-first-token response for long-session agents.

Source ↗

OpenAI Combines ChatGPT and Codex Usage Pools

Starting in December, usage across paid subscription tiers will be combined between ChatGPT and Codex.

OpenAI Launches Grid Protectors Program

Designed to enable future compute campuses to reduce power demand when the grid is strained, starting with the first site in Georgia.

Source ↗

OpenAI Deploys Activation Classifiers to Monitor Frontier Models

Starting from GPT-5.6, internal activations are used to score in real-time, blocking policy-violating responses and tool calls midway.

Source ↗